Abstract
Abstract (English version)
Verbal deception detection refers to techniques for identifying deceptive intent or content in written or transcribed statements. Despite decades of research, it remains a well-known and unresolved problem with significant implications in various high-stakes contexts, including criminal investigations, financial fraud, and deceptive behavior on online platforms. This thesis investigates the extent to which computational methods from artificial intelligence (AI) can be
employed to automate the detection of verbal deception, exploring both their opportunities and challenges.
We show that opportunities lie in the use of natural language processing techniques for the automated coding of statements and in the application of machine learning models for deception prediction.
However, because deception detection is a sensitive task, several challenges need to be considered. These include the difficulty of developing models that can generalize across different contexts and types of deception, as well as of developing models that are highly accurate, yet transparent and interpretable. Furthermore, given the possibility that automated methods may be integrated into real-world settings to support experts, more research is needed to understand the extent to which human users adopt AI-based judgments. Our experiments, in fact, show that people tend to remain somewhat skeptical of AI predictions, especially when they predict deception with high confidence.
Finally, we found that human oversight of AI predictions does not necessarily improve deception-detection performance, challenging the idea that humans and AI models can work together easily.
Overall, the findings of this dissertation highlight not only the contexts in which computational methods clearly outperform human performance but also their current methodological, conceptual, and practical limitations, underscoring the conditions under which their application in real-world settings remains constrained.
—
Verbal deception detection refers to techniques for identifying deceptive intent or content in written or transcribed statements. Despite decades of research, it remains a well-known and unresolved problem with significant implications in various high-stakes contexts, including criminal investigations, financial fraud, and deceptive behavior on online platforms. This thesis investigates the extent to which computational methods from artificial intelligence (AI) can be
employed to automate the detection of verbal deception, exploring both their opportunities and challenges.
We show that opportunities lie in the use of natural language processing techniques for the automated coding of statements and in the application of machine learning models for deception prediction.
However, because deception detection is a sensitive task, several challenges need to be considered. These include the difficulty of developing models that can generalize across different contexts and types of deception, as well as of developing models that are highly accurate, yet transparent and interpretable. Furthermore, given the possibility that automated methods may be integrated into real-world settings to support experts, more research is needed to understand the extent to which human users adopt AI-based judgments. Our experiments, in fact, show that people tend to remain somewhat skeptical of AI predictions, especially when they predict deception with high confidence.
Finally, we found that human oversight of AI predictions does not necessarily improve deception-detection performance, challenging the idea that humans and AI models can work together easily.
Overall, the findings of this dissertation highlight not only the contexts in which computational methods clearly outperform human performance but also their current methodological, conceptual, and practical limitations, underscoring the conditions under which their application in real-world settings remains constrained.
—
Abstract (Dutch version)
Het detecteren van verbale misleiding verwijst naar technieken voor het identificeren van misleidende bedoelingen of inhoud in geschreven of getranscribeerde uitspraken. Ondanks tientallen jaren van onderzoek blijft dit een bekend en onopgelost probleem met aanzienlijke gevolgen in diverse situaties met hoge inzet, waaronder strafrechtelijke onderzoeken, financiële fraude en misleidend gedrag op onlineplatforms. Dit proefschrift onderzoekt in
hoeverre computationele methoden uit de kunstmatige intelligentie (AI) kunnen worden ingezet om de detectie van verbale misleiding te automatiseren, waarbij zowel de kansen als de uitdagingen worden verkend.
We tonen aan dat kansen liggen in het gebruik van natuurlijke taalverwerkingstechnieken voor de geautomatiseerde codering van verklaringen en in de toepassing van machine learningmodellen voor het voorspellen van misleiding.
Omdat het opsporen van misleiding echter een gevoelige taak is, moeten verschillende uitdagingen in overweging worden genomen. Deze omvatten de moeilijkheid om modellen te ontwikkelen die kunnen generaliseren over verschillende contexten en soorten misleiding heen, evenals het ontwikkelen van modellen die zeer nauwkeurig, maar toch transparant en interpreteerbaar zijn. Bovendien is, gezien de mogelijkheid dat geautomatiseerde methoden in
de praktijk kunnen worden geïntegreerd om experts te ondersteunen, meer onderzoek nodig om te begrijpen in hoeverre menselijke gebruikers op AI gebaseerde oordelen overnemen. Onze experimenten tonen namelijk zelfs aan dat mensen de neiging hebben enigszins sceptisch te blijven ten aanzien van AI-voorspellingen, vooral wanneer deze met een hoge mate van zekerheid bedrog voorspellen. Ten slotte hebben we vastgesteld dat menselijk toezicht op AI-voorspellingen de prestaties op het gebied van bedrogdetectie niet noodzakelijkerwijs verbetert, wat het idee dat mensen en AI-modellen gemakkelijk kunnen samenwerken, ter discussie stelt.
Over het geheel genomen benadrukken de bevindingen van dit proefschrift niet alleen de contexten waarin computationele methoden duidelijk beter presteren dan mensen, maar ook hun huidige methodologische, conceptuele en praktische beperkingen, waarmee de omstandigheden worden benadrukt waaronder de toepassing ervan in de praktijk beperkt blijft.
| Original language | English |
|---|---|
| Qualification | Doctor of Philosophy |
| Awarding Institution |
|
| Supervisors/Advisors |
|
| Award date | 5 Jun 2026 |
| DOIs | |
| Publication status | Published - 2026 |
Fingerprint
Dive into the research topics of 'Opportunities and challenges of automated verbal deception detection'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver